Correct usage: no task-instruction prefix (matches how the model was trained and evaluated)

README.md

@@ -58,23 +58,33 @@ from sentence_transformers import SentenceTransformer
 
 model = SentenceTransformer("DataScience-UIBK/OBLIQ-IR-3B", trust_remote_code=True)
 
-# A per-task instruction prefix is prepended to the QUERY only.
-INSTRUCTIONS = {
-  "writing":  "Given a prose snippet, retrieve other snippets by the same author across different topics, based on stylistic fingerprint.",
-  "math":     "Given an example problem, retrieve other math problems whose solutions share the same abstract reasoning technique.",
-  "twitter":  "Retrieve tweets that implicitly express the latent stance described in the query, without surface mention of the topic.",
-  "congress": "Retrieve the specific Congressional hearing passage matching the user's lossy recollection of the exchange.",
-}
+queries   = ["A passage written with the same restless, aphoristic voice as this one."]
+documents = ["...your corpus documents..."]
 
-q = model.encode(["query: " + INSTRUCTIONS["writing"] + " " + "Your query text here."])
-d = model.encode(["passage: " + "Your document text here."])
+q = model.encode(["query: "   + x for x in queries])
+d = model.encode(["passage: " + x for x in documents])
 
-# rank by cosine similarity
 scores = model.similarity(q, d)
 ```
 
-**The `query: ` / `passage: ` prefixes are required** — the model was trained with them.
-Documents take no instruction prefix.
+**The `query: ` and `passage: ` prefixes are required** — the model was trained with them, and dropping
+them changes the embedding. Queries take no other prefix.
+
+> **Note on task instructions.** The training file carries a per-task instruction string alongside each
+> query, but the training script consumes only the query text, and the evaluation that produced every
+> number in the paper encodes the raw benchmark query with the `query: ` prefix alone. **Do not prepend a
+> task instruction** — the released checkpoint was neither trained nor evaluated with one, so adding one
+> moves you off-distribution. Task identity is carried by the mixture the model was trained on, not by a
+> runtime prefix.
+
+Truncate documents at 1024 tokens and queries at 256 to match the paper's evaluation:
+
+```python
+model.max_seq_length = 1024   # documents
+d = model.encode(["passage: " + x for x in documents])
+model.max_seq_length = 256    # queries
+q = model.encode(["query: "   + x for x in queries])
+```
 
 ## How it was trained